Detecting Edit Failures In Large Language Models: An Improved Specificity Benchmark

TL;DR

This paper introduces CounterFact+, an enhanced benchmark with a dynamic component and a KL divergence-based metric to better evaluate the specificity of model editing techniques in large language models, revealing their limitations.

Contribution

The paper extends the CounterFact benchmark with a dynamic component and proposes a KL divergence-based metric for improved specificity evaluation.

Findings

Model editing techniques show low specificity in evaluations.

Existing benchmarks fail to detect significant side effects.

CounterFact+ provides a more comprehensive assessment of model edits.

Abstract

Recent model editing techniques promise to mitigate the problem of memorizing false or outdated associations during LLM training. However, we show that these techniques can introduce large unwanted side effects which are not detected by existing specificity benchmarks. We extend the existing CounterFact benchmark to include a dynamic component and dub our benchmark CounterFact+. Additionally, we extend the metrics used for measuring specificity by a principled KL divergence-based metric. We use this improved benchmark to evaluate recent model editing techniques and find that they suffer from low specificity. Our findings highlight the need for improved specificity benchmarks that identify and prevent unwanted side effects.